Imagine writing a machine learning pokemon bot and having it metagame itself for about a week, then battling it and it beating you with something like Level One Aron except it's something humans haven't been doing.In post 521, Psyche wrote:I seem to have a robust interest in figuring out what would be the most systematic way to study a game like this one - specifically to find a battle strategy for the video games that's competitive against real people
One idea I had was iterations of bots that execute different battle strategies - defined with some explicit decision tree - that maybe I test at a place like Showdown. While this seems a decent approach for *testing* strategies, it seems a poor approach for strategy discovery, limiting me to just what I can imagine. Ideally I could pursue some kind of analytic approach, but it's tough to consider where one might start on that. Maybe I'm just not ready yet. Another possibility is consulting with regular players and experts and so on. But that defeats the point, and locks strategy to conventional wisdom, though requiring enough detail to support bot implementation and getting quantitative feedback through bot performance could be a decent corrective.
There's a way cognitive scientists come up with computational models of decision processes and maybe if I just get around to finding out how they do that, I'll have a better idea of what can be done here.
I'd go white as a sheet.
Oh it looks like you're just talking about the bot piloting a team that you already made for it.
That shouldn't be too hard.